@nexus-cortex/cli 4.124.36 → 4.124.38

package/.env.defaults

@@ -949,6 +949,45 @@ CORTEX_ENDTURN_GATE=true
 # unchanged low | medium | high) | legacy: the pre-R153c steps. reasoning_exhaustion events now carry {ladder, provider}.
 # [LEDGER 2026-09-30 @4.124.34] ON (bug fix) — unit-tested (effortLevers.test.ts). NO bench arm.
 #CORTEX_REASONING_EXHAUST_LADDER=
+# CORTEX_WALL_DROP — HB-WALL-DROP (DARK, 2026-10-01). c22: 69 reasoning-exhaustion walls in ~19.6K calls; each walled turn's ~65K of
+# reasoning stayed in history and rode along in every later call (photonic r2 +196K after three walls) and the retry re-read it uncached.
+# on: the walled assistant turn (reasoning only, no text, no tool call) is removed from the IN-MEMORY history before the retry (the session
+# file keeps it) and the nudge is carried at the tail of the newest tool_result (HB-TURN-STATUS carrier) → the retry is the exact pre-wall
+# prefix + the nudge. A turn with a tool call is never dropped. No tool_result before the wall → the nudge is pushed as a user message.
+# Event wall_drop {iteration, droppedChars, carrier, streaming}. Pairs with CORTEX_REASONING_EXHAUST_BACKOFF=false: post-wall recovery acts
+# ~90% at high OR low (t4cw 17/19 at high vs c22 59/64 at low) while the low step costs a mostly-uncached re-read (effort is in DeepSeek's
+# cache key: post-wall low call hit 0.25 vs high 0.77).
+# [LEDGER 2026-10-01 @next] DARK — BUILT + tested (wallGuard.test.ts 6/6; wallGuard.integration.test.ts sendMessage: retry carries no
+# walled turn + nudge on the tool_result tail; off = pre-existing; streamMessage wall cases skipped — scripted stream has no reasoning deltas).
+#CORTEX_WALL_DROP=
+# CORTEX_OUTPUT_CAP_TOKENS — HB-OUTPUT-CAP (DARK, 2026-10-01). Per-call output cap for the action model (fills maxTokens only when the request
+# did not set one; helper/mentor calls unaffected). c22 acting calls' reasoning: p99 14K, p99.9 40K, 7 of 19,337 above 48K → a wall costs
+# ~48K instead of 65,536 output tokens and minutes of decode. Recommended 48000. Not part of the prompt, so expected cache-neutral (verify).
+# [LEDGER 2026-10-01 @next] DARK — BUILT + tested (both loops: every action request carries the cap).
+#CORTEX_OUTPUT_CAP_TOKENS=
+# CORTEX_WALL_SUMMARY / _TIMEOUT_MS — HB-WALL-SUMMARY (DARK, 2026-10-01). On a reasoning-exhaustion wall, before HB-WALL-DROP removes the
+# turn, one helper call (the EndTurn judge's mentor wire, ~500-token budget) condenses the walled reasoning (head 4K + tail 16K chars) into
+# CONCLUDED / STUCK ON / NEXT; the result rides INSIDE the wall nudge (same cache-safe tail), attributed as written by the harness — item-11c
+# FOREIGN-THINKING doctrine (never the thinking channel). Fail-open: off / no reasoning / timeout (_TIMEOUT_MS, default 30000) / error → the
+# plain nudge. Event wall_summary {iteration, inputChars, summaryChars, ok, ms, streaming}. Aim: break the overthinking spiral without the
+# effort backoff (which busts the cache) — pair with CORTEX_WALL_DROP=on + CORTEX_REASONING_EXHAUST_BACKOFF=false.
+# [LEDGER 2026-10-01 @next] DARK — BUILT + tested (wallGuard.test.ts helpers; integration sendMessage: summary in the nudge, helper error →
+# plain nudge). NO bench arm yet.
+#CORTEX_WALL_SUMMARY=
+#CORTEX_WALL_SUMMARY_TIMEOUT_MS=
+# CORTEX_KILL_GUARD — HB-KILL-GUARD (2026-09-30, DEFAULT ON). The Bash tool runs every command as `bash -c "<command>"`, so a model's
+# `pkill -f "<pattern>"` always matches its OWN shell, and `pkill -f node` / `killall node` / `kill $(pgrep -f node)` reaches the harness
+# server. TB4.0 c21ba cumulative-layout-shift r3 (4.124.36): `pkill -f "next-server"; pkill -f "next start"; ...` → the tool shell got
+# SIGTERM and the run ended NonZeroAgentExitCodeError 29 min in. on (default): pkill / killall / pgrep wrappers first on the Bash tool's
+# PATH (foreground + run_in_background spawns); targets resolved with the real pgrep, then PROTECTED pids dropped — the calling shell +
+# its ancestors, the harness server (CORTEX_SERVER_PID) + its ancestors, and every process descended from those ancestors other than
+# through the server (bench adapter request, process recorder); anything the model started stays killable. Skips are reported on stderr
+# ("[harness] pkill: left N protected process(es) alone"). Not covered: `kill <literal pid>`, `ps | xargs kill`, TmuxSession panes.
+# off: no wrappers, spawn env unchanged.
+# [LEDGER 2026-09-30 @next] ON (bug fix) — tested on live processes (executors killGuard.test.ts 7/7: self-match shell survives, harness
+# server + adapter sibling survive a matching pattern, killall spares the sibling, pgrep hides the harness, the c21ba command shape through
+# the real Bash tool); ShellTool/BashOutput/KillShell suites unchanged (145 tests).
+#CORTEX_KILL_GUARD=
 # (no lever) R154 HB-IMAGE-REJECTION-HEAL (4.109.0, 2026-09-16): retro-console-soc generated a 128x128 8-bit GRAYSCALE PNG, called
 # ReadImage, and DeepSeek answered the next continuation with `400 .messages[290].image[0]: You have uploaded an unsupported image`;
 # the exception escaped after the tool batch, the already-executed guard left the loop, and the R29a synthesis failed on the same block —

package/package.json

@@ -1,11 +1,11 @@
 {
   "name": "@nexus-cortex/cli",
-  "version": "4.124.36",
+  "version": "4.124.38",
   "description": "Nexus Cortex CLI - Terminal interface for multi-provider LLM orchestration",
   "main": "dist/index.js",
   "type": "module",
   "bin": {
-    "cortex": "./bin/cortex.js"
+    "cortex": "bin/cortex.js"
   },
   "scripts": {
     "build": "cp ../../.env.defaults ./.env.defaults && tsc",
@@ -19,7 +19,7 @@
     "prepack": "node ../../scripts/copy-pkg-cortex-scaffold.mjs"
   },
   "dependencies": {
-    "@nexus-cortex/core": "4.124.36",
+    "@nexus-cortex/core": "4.124.38",
     "chalk": "^4.1.2",
     "cli-spinners": "^2.9.0",
     "commander": "^11.0.0",